Slip Validation and Prediction

for
Mars Exploration Rovers
Jeng Yen
Jet Propulsion Laboratory,
Pasadena, California, USA
jeng.yen@jpl.nasa.gov

Abstract
This paper presents a novel technique to validate and predict the rover slips on Martian surface for NASA's
Mars Exploration Rovers. Different from the traditional approach, the proposed method uses the actual velocity
profile of the wheels and the digital elevation map (DEM) from the stereo images of the terrain to formulate
simplified equations of motion of the rovers. A weighted factor to the wheel-ground speed from the empirical
data comprises the velocity equations of the simplified differential-algebraic system of the rover motion.
Applying the discretization operator to these equations, the full kinematics state of the rover is then resolved by
the configuration kinematics solution in the Robot Sequencing and Visualization Program (RSVP) [1,6,10]. This
method produced accurate simulation of the rover movements compared with these of the earth testing vehicle.
Using the telemetry from the onboard Visual Odometry [2], the simulated rover path also compares well with the
actual track of the vehicle. Preliminary results indicated that the proposed computational method is very
effective in planning the path of the rovers on the high-slope areas.

Keywords: MER, Mars Rover, multi-body simulation, configuration kinematics, autonomous navigation.

rover. This requires effective modeling of the rover-

1 Introduction
In January 2004, NASA successfully landed two robot
geologists: Spirit and Opportunity Mars Exploration
Rovers (MER), on the surface of Mars. The mission's
primary objective is to find evidence of past water on
Gusev Crater and Meridiani Planum, two locations on
the opposite sides of the planet. To date, after 20
kilometer accumulated distance of driving, they have
succeded in the primary goal and are still roaming on
the Martian surface for scientific targets. This
tremendous acheivement owes to the ability of the
on-board mobility and the ground sequence
simulation for planning safe path for the rovers in the
harsh Martian surface. We will address the MER
vehicle simulation in this presentation.
Simulations of space-borne systems have been well-
developed and successfully applied to many of the
past and current NASA missions. The modeling and
simulation of spacecraft has been carried out using
multi-body system dynamics [4]. The design and
operation of the spacecraft are based on the predicted
behavior using high-fidelity simulation tools [5]. The
modeling of robotic vehicles for the surface operation
is, however, very different from those of traditional
spacecraft operations. The most important aspect of
the rover's model is the need to interact with the
surrounding terrain. Based on the knowledge of the
terrain, the simulator will predict the states of the
terrain interaction. In addition, the multi-body rover
model should include all the motorized mechanisms
that are commandable for a comprehensive sequence
simulation.
Robot Sequencing and Visualization Program (RSVP)
provides the capability to operate and control the
MER rovers. Typically, the rovers are commanded
once per Martian day, called a sol. A sequence of
commands sent in the morning specifies the sol's
activities: what images and data to collect, how to
position the robotic arm, or where to drive. At the
end of each sol, the rovers send back the data and
images human operators will use to plan the next sol's
activities. Using the command-level editing and the
sequence-level simulation of RSVP, the operators can
select the next sol's mobility commands and visualize
the predicted motion [6,8].
The sequence rehearsal tool in RSVP is based on
modeling and simulation of the multi-body
mechanical system [10]. The methodology has been
developed to support a real-time interactive graphics
mode for the visualization tool, using the
configuration kinematics algorithm [3] and 3D terrain
models. The sequence simulation is carried out using
the on-board flight software modules for realistic
rover behavior.
Determined by the solution of six wheel-terrain estimates, and has become a critical component of
contact equations, the algorithm solves the vehicle's rover's safety systems. The telemetry of these
wheels, steering and suspension linkages, and the onboard positions will be used to validate the
position and orientation of the chassis. It treats the proposed slip model.
underlying mathematical model as an inverse-
Conventional simulation of wheeled vehicles often
kinematics problem, and carries out the solutions
uses a contact compliance formulation to deal with the
using the computational techniques for constrained
complex wheel-terrain surface contact model [4]. The
optImlzation. In this framework, the objective
contact forces on the wheel are represented by a linear
functions are comprised of three conditions: the
spring-damper actuator. Based on the penetrative
rover's internal differential mechanisms, the wheel-
distance between the terrain and the wheel, a force
terrain contact, and the commanded rover location and
will be applied to the vehicle's equations of motion to
heading as shown in [3].
generate the dynamical effects from the contact.
During the surface mission phase, the configuration However, the numerical solutions of this dynamical
kinematics simulation has provided fast and high system suffers from instability and because the rough
quality results for path planning on relatively level terrain profile and the physical limits of the linkages
ground. The simulation results of the rover position (e. g., the bumper-stops) can inject impulsive forces in
and orientation using terrian DEM from the aquired such systems. All these modeling and numerical
images are remarkablely close to those of the on- difficulties prevent a novel solution that achieves the
board estimation [8]. But on steep hillsides, in mixed real-time simulation of the rover traversal on a rocky
sand/rock terrains inside craters, and even when terrain.
crossing sandy ripples in the otherwise flat plains of
To overcome these difficulties a set of simplified
Meridiani, the simulation has not, as expected, been
dynamic equations is first developed, and then a
able to acurately predict the rover position due to
heuristic wheel-ground speed estimation is applied to
large slips. The only on-board sensor that can detect
reduce the contact dynamics to a psuedo-dynamics for
position slip is the Visual Odometry (VisOdom) [2].
the slip model. This technique has been implemented
The VisOdom software compares two pairs of images
in RSVP to compute ARC and TURN for MER
to detect and track a set of features between these
vehicles. In the following section, we'll present the
images. The motion of the features is used to update
underlying framework ofthe MER vehicle simulation.
the vehicle's on-board position estimate according to
the algorithm described in [9]. On high-slip areas,
VisOdom can produce accurate onboard position

Rover.

Low

Panc~r:n
• o· _
":"Calibr~i6n.~--~
~:... . ar~f:' _-'~-~"i:-~-:
2 Mars Rovers Simulation
In RSVP, the MER vehicle model is represented by a
set of hierarchical sub-graphs of the mechanism
models for the primary motion systems. These sub-
graphs are the foundation block for receiving sensed
data, interpreting commands and predicting the
physical states of the corresponding mechanical
systems as shown in Fig. 1. The mobility mechanism
consists of the Rocker-Bogey-Differential (RBD)
suspension and the four-wheel steering systems. As
shown in Figure 1, the suspension and the rover's
kinematics, e.g., position and attitude, comprise a
multi-body system of 10 states. These are the basic
states of the rover's model for mobility.
Note that six wheels with four steering motors
comprise additional states of the full vehicle system,
but these acuators are fully cotrolled by the onboard
mobility software. Thus, the steering and driving
motors are not accounted for by the RSVP rover
simulation model. In particular, the onboard mobility
software controls the drive motors following a
specific velocity profile to achieve Ackermann
steering on flat ground [9].
2.1 Constrained Multi-body System
The equations of motion of multi-body dynamic
systems can be written as
W= M-1[f - G T
A]
(1)
g(q) = 0
where q is the generalized state, iJ
and ?} are the
velocity and acceleration of the generalized states, M
is the generalized mass matrix, f is the applied force,
d
g is the algebraic constraint and G = - g is the
dq
T
Jacobian of the contraint, and G A represents the
constriant reaction force, where A is the Lagrange
multiplier. The generalized coordinates q of the
MER vehicles consist of 10 states, the vehicle's
in this approach often imposes a restricted step size of
the discretization methods because of the so-called
artificial stiffness. Alternatively, Eq. 1 can be
rewritten to a state-space form, e.g., a set of
differential equations representing the true degree of
freedom in the system. One of these reduction
techniques is the generalized coordinate partitioning
method [4] which has been widely used in the
simulation of ground vehicles in the automotive
industry.
2.2 Coordinate Partitioning for
Constrained Multi-body Dynamics
The state-space representation of the rover motion of
(1) can be expressed as
E= XT M-1[f - GTA] = fy
)if= yT M-1[f - GTA] = I;: (2)
t= rl M-1[f - G A] T
= Tz
where [x,y,e] are the rover x-coordinate, y-
coordinate and heading, the projection operators are
defined as x =XTq,y =yTq,e=rlq. It is
interesting to note that the closed form state-space
representation of Eq. 2 for a general multi-body
dynamical system can be very difficult to obtain. The
right-hand-side of Eq. 2 is often evaluated using the
full generalized acceleration in Eq. 1.
Applying the time integration operator to Eq. 2 yields
the independent coordinate of q. Applying the
solution of the independent coordinates, the
dependent coordinates in q can be obtained by
solving the constraint equations, e.g.,
x = Jv,dt JJ J,dt=
Y = f ydt If I;:dt
V =
(3)
61 = f wzdt If Tzdt
=

position and orientation states, joint angles of the left g(q) = 0

and right rockers, and joint angles of the bogeys. The
constraints g containts a linear equation of the where [vx,vy,wJ is the velocity of the generalized
differential, e.g., the joint angles of the left and right coordinates [x,y, 61].
rockers are equal and of the opposite sign, and six
rigid contact equations 2.2.1 Conf~guration Kinematics
geol1faet(q) = dist(w,c) = 0 Apply directly the desired rover position from an
ARC or a TURN command, Eq. 3 can be written as
where the distance of the wheel center to the terrain is
equal to the wheel radius. Thus, Eq. 1 represents a x - x are = 0
system of only three degrees of freedom.
y - Yare = 0
The solution of Eq. 1 can be obtained using the (4)
standard methods for Differential-Algebraic 61- eare = 0
Equations (DAE) [12], where a class of the g(q) = 0
discretization operators is directly applied to the
differential part of Eq. 1. The stability of the solution
where [X arc 'Yare' ()arc] is the result of commended
location and heading. The solution of Eq. 4 is then
obtained by a Newton-type iterative method. Regular
Newton-type iteration requires that g is smooth (e.g.
2
gEe has 2 nd order derivatives) to ensure a fast
convergence. This prerequisite of a robust
convergence is violated since the roughness of the
terrain has been embedded in the contact equations.
When the rocker or bogey linkages reach their limits,
the abruptness of the iteration can induce
unpredictable solutions of the configuration. To
overcome these numerical difficulties, we apply a
weight factor to the residual of each contact equation.
During the iterations, the weight factor for a given
wheel can be reduced to zero to relax the contact
condition. Whenever the wheel leaves the ground, its
corresponding weight factor is set to zero for a total
relaxation of this wheel-terrain contact. The re-
scaling of the weight factors is coupled with the
global search algorithm, which can detect the joint
limits associated with each wheel-terrain contacts, and
can sample small perturbations around the contacting
locations to determine the occurrence of a separation
of the wheel and the ground.
The step-selection strategy used in the global search is
a backtracking line search algorithm that monitors the
progress of the iteration. For a smooth terrain profile,
the iterative solution generated by the Newton method
converges very rapidly to a local minimum of the
nonlinear equations. However, the rate of
convergence can be tremendously decreased when a
non-smooth terrain profile appears. Special care is
taken to maintain a robust and efficient solution in the
case of a non-smooth terrain profile. Although the
problem in hand is ill posed (i.e., it is well-known that
the Newton method cannot treat non-smooth
equations), we developed a heuristic solution to ease
the computational difficulty in the iterations. In
practice, the wheel-terrain contact is treated as a non-
penetrative type, which is not a realistic portrait of the
nature of the wheel-terrain interaction. Therefore, a
search direction to the wheel-terrain contact may not
be in-line with the normal direction of the terrain (at
the contact location), instead; it could be anywhere
along the perimeter of the wheel. The heuristic leads
to modeling the wheel-terrain contact equation as the
distance constraint between the wheel center and the
terrain profile. On level ground, the solution of Eq. 3
by the aforemention techniques results in an efficient
and accurate estimation of the commanded location of
the MER vehicles [6,10].
2.2.2 Estimation of Independent Velocity
Equation 3 can be again re-written as a first-order
differential-algebraic equations by applying the time
differentiation operator to the state variables [x,y, ()]
ofEq.4. This yields
N-v x =0
y,V-vy =0
(5).
t-OJ =0
z
g(q) = 0
In fact, Eq. 5 can also be obtain from Eq. 4 if we
declare that the time derivatives are
d d d
Vy = dt x are' Vy = dt Yare' and OJz = dt ()arc
The vehicle's speed when performing ARCs and
TURNs has been measured in the earth-testbed, so
these empirical vehicle speed profiles/tables may be
used directly in Eq. 5. A straightforward calculation
leads to the "slip tables" approach which use only the
empirical data to approximate the slips. Then the
modified Eq. 3 is resolved for the add-on slips.
However, only the gross motion can be represented by
these tables, and the dimension of these tables can be
very large.
Using rigid-body motion, the velocity of the
independent states can be obtained from a consistent
wheel-ground speed. Since the onboard drive motor
controller is always keeping up with a pre-defined
profile, we can use the speed profile and the local
terrain geometry for an approximated velocity of the
independent variables. Moreover, the wheel-terrain
slip can be incorporated in this approximation so that
the vehicle total slip can be computed accordingly. A
heuristic approach is to "avarage" the wheel-ground
speed of each wheel for the independent velocity
[vx,vy,OJJ. Let the i-thwheel's velocity projected
onto the plane interpolation of the terrain patch under
the wheel be
(6)
where (1- p;p;T) is the projection and v;w is the
nominal wheel speed. Assuming the rigid contact
condition, the wheel-ground speed is along the
tangential direction of the contact plane, as shown in
Eq. 6. For all six wheels in contact with ground, the
rover velocity is a linear combination of the wheel-
ground velocity, e.g., vehicle's velocity is within the
convex hull of all the wheel velocity. Apply the
partitioning operator and a weight factor to each of
the driving wheels yields
v. =XT"avw+d
X" ~ I f X
(7)
where a; and prepresent the weight factors, and 'i
is the location of the wheel in the rover's navigation
frame, and (d"dy ) is the sliding vector of the rover.
Typically, the sliding vector is along the down slope
direction on a local patch of the terrain where the
vehicle is treated as a point-mass on a gravitational
field, e.g.,
dx = JJ .fxgdt
I

dy = ff rgdt
.f y

where frg and .1"/ are the gravitational forces acting

on the independent coordinates (x,y), respectively.
Using Eqs. 5 and 7, the rover position and orientation, Figure 3 Up Slope Driving 14-16% Slip
and the joint angles of the rockers and bogeys can be
obtained. The weight factor is selected based on the
local geometry, e.g., the wheel-ground angle. The
magnitude of sliding vector is determined by the
experimental data of the slip tables. Since the
maximum tilt of the static stable configuration of the
MER vehicles is 35 degrees, the weight factor and
sliding factor are obtained from the results
interpolated between 0 to 30 degrees.

3 Numerical Examples ~ '" .~ ..

The numerical test for the rover slip model is

illustrated on the Figs. 2-5. The rover started on a flat
surface with 14 degrees slope at the origin (0,0). The
testing sequence comprised of a 3-meter ARC driving Figure 4 Side Slope Driving 10% Slip
downhill, then a 3-meter ARC back up to the origin,
then TURN to 0 degree heading, then driving back
and forth with two 3-meter ARCs. Comparing the a ...--..... ~III' "'u. I II ..... ' I,,· I I ~ Ih~ :r "!'"

"no slip" simulation using Eq. 4, indicated by the

solid rover in the Figs., with the simulation using Eq.
5 and 7, indicated by the watermarked rover, the
slippage predicted by the simulation matches well
with those of the experiments. When driving in the
downhill direction, Fig.l shows a 14% down-slope
slip of the straight arc. The uphill straight arc
illustrated in Fig. 2 shows a similar down-slope slide.
After a tum to line up to a side slope direction, the
forward and back arcs slide toward the down-slope
direction at about 10 % of the traveling distance,
/
Figure 5 Side Slope Driving 10% Slip
." :,,~ ~~""'. '.Iv~\·:·~": :.'~I<'<'.': 11"<".<"

We learned during our initial drives in each terrain

that driving on level ground typically leads to accurate
and predictable mobility performance; e.g., Spirit only
accumulated 3% position error over 2 kilometers of
driving [8]. Before driving into Endurance Crater, a
series of driving tests were carried out to establish a
set of slip tables for the MER vehicles. These tables
measure the slippage percentage based on the tilt
angle of the vehicle assuming a flat surface
,
i, r~~~ ',I~~~ ',II~ ,)r~II~.' underneath. We have implemented the slip tables in
RSVP simulation as the baseline method to compute
the weight factors in Eq. 7. From the test results, the
Figure 2 Down Slope Driving 14% Slip vehicle exhibits about 1% slip (of the total driving
distance) per degree on a slope ranging 10-20 degrees
sandy surface. Similar performance is seen on the
MER vehicles when driving on the Martian surface
covered with "blue berries", which are small round
pebbles.
3.1 High Slope Hillside
While most of the distance covered by the rovers was
on level ground, most of the sols and most of the
approach drives occurred on slopes. The rovers
invariably slip when driving on slopes, making
VisOdom essential for safe and accurate driving. To
construct a successful drive on high slope areas,
accurate estimation of the commanded location can
only be obtained with the correct vehicle slips.
On sol 1211, Opportunity drove toward the rim of
Victoria crater to take images of Cape St. Mary. The
vehicle was on a modest terrain of 7-10 degree slope
tilted left as shown in Fig. 6, but the drive would take
the vehicle very close to the edge of the Crater ( about
2 meters from the visible rim side), making the drive
nerve-racking. We took the vehicle 4 meters forward
and pointed the camera to Cape St. Mary 100 degrees
left and 70 meter away.
*.-c,.
.:.~ ~
i Integrating textual and visual command interfaces for
• '.
Figure 6 Approach to Victoria Rim with 5% Slip
Three rover paths in the blue lines are shown in Fig.
6, where the left one is the obtained by the CK
method, the right line is the prediction of proposed
method. The Onboard Visodom correction of the
rover location is the blue line in the middle. Note that
the 5% slips along the right side of the slope matched
well with the VisOdom data.
4 Acknowledgements
Thanks to Jeff Biesiadecki for insightful discussion on
the motor controller onboard the MER flight system,
RSVP team lead Brian Cooper and the team members
(Frank Hartman, Scott Maxwell, and John Wright) for
supporting the inclusion of the slip model in RSVP,
Mark Maimone for fruitful discussion on validatiing
the slip model using the telemetry and VisOdom data.
The work described in this paper was carried out at
the Jet Propulsion Laboratory, California Institute of
Technology, under a contract to the National
Aeronautics and Space Administration.
References
[1] B. Cooper, F. Hartman, S. Maxwell, and l
Wright, l Yen. "Using RSVP for Analysing State and
Previous Activities for the Mars Exploration Rovers."
In SpaceOps, Montreal, Canada, 2004.
[2] Yang Cheng, Mark Maimone, and Larry
Matthies. "Visual odometry on the Mars Exploration
Rovers." In IEEE Conference on Systems, Man and
Cybernetics, The Big Island, Hawaii, USA, October
2005.
[3] l Yen, A. Jain and B. Balaram, "ROAMS: Rover
Analysis, Modeling and Simulation," in the proc. Of
5th iSAIRAS, ESA, Netherland, 1999.
[4] E. J. Haug, "Computer Aided Kinematics and
Dynamics of Mechanical Systems", Allyn-Bacon,
1989.
[5] Biesiadecki, l, Henriquez, D., and Jain, A., "A
Reusable, Real-Time Spacecraft Dynamics
Simulator," Proceedings of the Sixth Digital Avionics
Systems Conference, Irvine, CA, October 1997.
[6] S. Maxwell, B. Cooper, F. Hartman, J. Wright, l
Yen, and C. Leger, "The best of both worlds:
mars rover operations," in Proc. IEEE Syst., Man,
Cybern. Conf, 2005, vol. 2, pp. 1384-1388.
[7] Rongxing Li, Brent A. Archinal, Raymond E.
Arvidson, Jim Bell, Philip Christensen, Larry
Crumpler, David l Des Marais, Kaichang Di, Tom
Duxbury, Matt Golombek, John Grant, Ronald
Greeley, Joe Guinn, Andrew Johnson, Randolph L.
Kirk, Mark Maimone, Larry H. Matthies, Mike Malin,
Tim Parker, Mike Sims, Shane Thompson, Steven W.
Squyres, and Larry A. Soderblom. Spirit rover
localization and topographic mapping at the landing
site of Gusev Crater, Mars. JGR-Planets, Special Issue
on Spirit Rover, 111(E02S06, doi: IO.I029/2005JE002483).
http://www .agu.org/j ournals/ss/SPIRIT 1/.
[8] Mishkin, A.H.; Limonadi, D.; Laubach, S.L.;
Bass, D.S.; "Working the Martian night shift - the
MER surface operations process," Robotics &
Automation Magazine, IEEE Volume 13, Issue 2,
June 2006 Page(s):46 - 53.
[9] Biesiadecki, l l, Leger, P. c., and Maimone, M.
W. (2007). Tradeoffs between directed and
autonomous driving on the Mars Exploration Rovers.
International Journal of Robotics Research, 26(1):91-
104.
[10] J. Yen, B. Cooper, F. Hartman, S. Maxwell, and
J. Wright. Sequence rehearsal and validation on
surface operations of the Mars Exploration Rovers. In
SpaceOps, Montreal, Canada, 2004.
End of File